A.2 Evaluation of Bayesian Cognitive Models in the Prediction of Human Behavior

- Tim Schürmann –

When dealing with privacy-relevant decision making in the usage of online services, human behavior has been termed irrational in past research. The tendency to behave contradictory to one’s stated worries about privacy violations has been established as the privacy paradox. The reasoning behind said paradox involves claims about bounded or limited capabilities of human subjects, which in turn are expected to lead to less than rational observed behavior. Deviance from statistically optimal behavior has been observed in a wide range of psychological experiments, with phenomena like probability weighting and matching being raised as indicators of human subjects not adhering to the concept of Bayesian rationality. However, implementations of rational process models like the sampling hypothesis generally show that a local lack of adherence to Bayesian rationality can turn out to be globally rational. Drawing from an example of which traffic route to pick in a car ride, they show that drawing very few samples from a posterior will lead to global utility maximization.

Transferring the concept of the sampling hypothesis to privacy-related decision making, I aim to explain deviations from a Bayesian posterior by comparing subject behavior to a sampling-based Bayesian agent. Subjects will be asked to decide whether or not to install a smartphone app, expecting their decision to be a function of sampling from their Bayesian posterior distribution after considering individual utility the app might provide for them. The probability of deciding to install the app is expected to follow a Bernoulli distribution defined by the sampling process. This allows an investigation of the expected amount of deviance from an individual’s optimal choice, describing local seemingly irrational behavior as a consequence of global rationality. Assuming that the sampling-based agent is able to replicate subject behavior, I will then draw conclusions on current means of psychological interventions designed to empower online privacy protection, how they might affect the decision making process of a samplingbased agent, and how to take advantage of the rational process model characteristics to increase the chance of privacy-protecting behavior in human subjects.

Tandem partner: A.1